Rede Neural Convolucional na Reconstrução de Pulsos em Alta Taxa

Precise pulse reconstruction in high-rate environments is a critical challenge for high-energy physics calorimetry systems, where signal pile-up and baseline fluctuations cause severe distortions. This paper presents a streamlined Convolutional Neural Network (CNN) architecture for real-time energy estimation in the ATLAS Tile Calorimeter channels under HL-LHC operating conditions. Developed under strict latency and hardware resource constraints, the proposed topology requires fewer than 160 adjustable parameters, comprising two 1D convolutional layers and one output neuron with ReLU activation. The model was evaluated with simulated data across low, medium, and high channel occupancy levels. Results show that the CNN outperforms the conventional Wiener filter in all scenarios, proving highly robust against baseline shifts driven by the long pulse tail and LHC bunch gaps. The low computational complexity of the proposed algorithm ensures its feasibility for experimental integration and high-speed processing on FPGA or GPU platforms.